An On-line Decision-Theoretic Golog Interpreter

We consider an on-line decision-theoretic interpreter and incremental execution of Golog programs. This new interpreter is intended to overcome some limitations of the off-line interpreter proposed in [Boutilier et al., 2000]. We introduce two new search control operators that can be mentioned in Golog programs: the on-line interpreter takes advantage of one of them to save computational efforts. In addition to sensing actions designed to identify outcomes of stochastic actions, we consider a new representation for sensing actions that may return both binary and real valued data at the run time. Programmers may use sensing actions explicitly in Golog programs whenever results of sensing are required to evaluate tests. The representation for sensing actions that we introduce allows the use of regression, a computationally efficient mechanism for evaluation of tests. We describe an implementation of the on-line incremental decision-theoretic Golog interpreter in Prolog. The implementation was successfully tested on the B21 robot manufactured by RWI.